Can AI Uncrumple This? Finding a Hidden Buckle
An I-beam buckled by a hidden nine-number warp. A five-stage net chain reads 64 dots and rebuilds the nine numbers step by step. Held-out error 0.075 to 0.043, hidden buckle back to 0.014.
Overview
An I-beam comes in buckled past recall and the buckle is the puzzle. Input is fixed: a prismatic I-section beam (120 long, 50 deep, 36 wide flanges, 8 thick flanges on a 6 web, 300 swept vertices), bent by a hidden warp of nine numbers. Only 64 surface dots are measured. Press Run and a small fixed chain reads the dots and rebuilds the nine numbers in five steps. Held-out shape error falls 0.075 to 0.043. The one beam morphs from gray (bent) to amber (restored) in place.
Bars show warp numbers: gray first guess, amber final. Side bars show per-number travel. A log curve shows mismatch by step. Each shown state comes from nets. No optimizer runs in replay.
Run plays in chunks with 450 ms per stage plus a 300-frame unwind on a log-time path. Reveal eases the beam from the buckled state toward the first guess over 48 frames while its tint warms from gray.
Why Backwards Is Hard
Forward is cheap: known buckle gives shape by flow. Backwards is hard in kind, not just in work. A fold map kills facts: two rest points land on one seen dot and no smooth flow fits, at any size. So this page draws a firm line: each learnable warp is a smooth buckle flow, invertible by backward integration. The one shown failure sits on display on purpose. Each smooth fit turns down a mirrored probe, whose mismatch sticks at 0.19 against the mystery’s 0.014.
The net never learns part names. It inverts buckle numbers against a known beam shape. That is why tens of thousands of weights do where sight would need millions.
Some help comes from known context: known beam shape, mean and scale stats, ID-checked mesh pair. “No help” would overstate.
Warp Field
The hidden buckle is a still buckle-styled speed field grown over unit time: nine fixed modes (five riding the beam length as bow, sag, and ripple, four across the section) with size numbers. Flow of a smooth field cannot fold, so invert holds by build. Tests pin round trips to 1e-4 and match a separate numpy build to 1e-9. The map bends non-linearly in its numbers (ODE sense). With ordered dots and a linear map this would be plain least squares and the net would be stagecraft. The buckle is what earns the net.
Chain
One net cannot sculpt fine enough. A single pass lands near but stays wide. So training is greedy in stages: stage zero maps 192 scaled dots to nine warp numbers, and each next sculptor reads dots plus running guess plus current mismatch and returns a fix. Four (64,32) sculptors follow stage zero. About 144k weights in all (143,917). Each stage locks before the next trains. Each stage edge is thus a real testable sculpt step, and the widget replays those beats.
All training runs offline in the uncrumple repo with fixed seeds: 1,500 warped pairs plus 256 held-out, held-out gates per stage, and a run kept only when each stage beats the last on held-out rebuilt shapes. Weights ship fixed. The widget only runs forward passes. Nets sculpt. Physics only scores.
Note on numbers: 0.075 to 0.043 is held-out shape error from training. The widget curve shows dot-space mismatch per stage: 0.0646, 0.0335, 0.0219, 0.0193, 0.0140 on a log scale. They track related but distinct gaps. Hidden buckle returns to 0.014 rebuilt-shape RMS (0.0140 measured).
Checks
Core gates pin exact identity map, round trips to 1e-4, clear Jacobians on large warps, converged growth across grids, numpy match to 1e-9, exact I-section area 780 with closed outward volume 93,600, 64 deterministic stride samples, forward pass pinned to hand math, hidden shape clearly nearer rest at close with each beat inside reach, rebuilt shape within 0.014 of rest, bit-same replay with no random bits, step twins matched to full runs, weights paired to mesh by ID, mirrored fold probe turned down while the puzzle closes, in-frame charts. The widget holds its own full recovery run. The trainer holds its own suite: mirror truth, set fixity, stage shapes, and a mono gate that blocks writes when a stage slips.
Honest Limits
A shape gallery, not a metals lab, and only in spring-back range. Forward math is buckle-flow math, not stock math: no stress, no yield, no history. The net takes no stock inputs at all (no stiffness, no harden, no rub). What it inverts is the shape part of reversible buckle, the same class as spring-back: one rest shape per buckled shape by build. Yield would kill just the facts the inverse needs. Many rests and paths map to same buckled parts, so yielded buckle and true self-touch fold stay out of range by physics, not by work. A new shape needs retrain: mesh and weights move as an ID-checked pair.
The 3.2 MB weight file is the heaviest file on the site and loads lazily with checks before run. Training code lives in uncrumple. This page ships its output as one fixed JSON file (public/data/uncrumple-weights.json) and links back to the rig that made it.
References
- diolei. n.d. “uncrumple — Staged Warp Trainer.” GitHub repository. Accessed October 5, 2026. https://github.com/diolei/uncrumple
- Dollar, Piotr, Peter Welinder, and Pietro Perona. 2010. “Cascaded Pose Regression.” In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 1078–1085. https://doi.org/10.1109/CVPR.2010.5540094